getdocf258. 308KB Jun 04 2011 12:05:11 AM

J
ni
o
r
t
Elec

o

u

a
rn l

o

f
P

c

r

ob
abil
ity

Vol. 10 (2005), Paper no. 2, pages 21-60.
Journal URL
http://www.math.washington.edu/∼ejpecp/
Convergence of Coalescing Nonsimple Random Walks to the Brownian Web
C. M. Newman1 , K. Ravishankar2 and Rongfeng Sun3
Abstract
The Brownian Web (BW) is a family of coalescing Brownian motions starting from every point in space and time R×R. It was first introduced by Arratia, and later analyzed in detail by Tóth and Werner.
More recently, Fontes, Isopi, Newman and Ravishankar (FINR) gave a
characterization of the BW, and general convergence criteria allowing
in principle either crossing or noncrossing paths, which they verified
for coalescing simple random walks. Later Ferrari, Fontes, and Wu
verified these criteria for a two dimensional Poisson Tree. In both
cases, the paths are noncrossing. To date, the general convergence criteria of FINR have not been verified for any case with crossing paths,
which appears to be significantly more difficult than the noncrossing

paths case. Accordingly, in this paper, we formulate new convergence
criteria for the crossing paths case, and verify them for non-simple coalescing random walks satisfying a finite fifth moment condition. This
is the first time that convergence to the BW has been proved for models with crossing paths. Several corollaries are presented, including an
analysis of the scaling limit of voter model interfaces that extends a
result of Cox and Durrett.
1

Research supported in part by NSF grant DMS-0104278
Research supported in part by NSF grant DMS-9803267
3
Research supported in part by NSF grant DMS-0102587

2

21

Keywords and Phrases: Brownian Web, Invariance Principle, Coalescing
Random Walks, Brownian Networks, Continuum Limit.
AMS 2000 Subject Classifications: 60K35, 60J65, 60F17, 82B41, 60D05
Submitted to EJP on November 2, 2004. Final version accepted on January

18, 2005.

22

1

Introduction and Results

The idea of the Brownian Web dates back to Arratia’s thesis [1] in 1979, in
which he constructed a process of coalescing Brownian motions starting from
every point in space R at time zero. In a later unpublished manuscript [2],
Arratia generalized this construction to a process of coalescing Brownian motions starting from every point in space and time R × R, which is essentially
what is now often called the Brownian Web (BW). He also defined a dual
family of backward coalescing Brownian motions equally distributed (after a
time reversal) with the BW which we call the Dual Brownian Web. Unfortunately, Arratia’s manuscript was incomplete and never published, and the
BW was not studied again until a paper by Tóth and Werner [27], in which
they constructed and analyzed versions of the Brownian Web and Dual Brownian Web in great detail and used them to construct a process they call the
True Self Repelling Motion.
In both Arratia’s and Tóth and Werner’s constructions of the BW, some
semicontinuity condition is imposed to guarantee a unique path starting from

every space-time point. More recently, Fontes, Isopi, Newman and Ravishankar [15, 16] gave a different formulation of the BW which provides a
more natural setting for weak convergence, and they coined the term Brownian Web. Instead of imposing semicontinuity conditions, multiple paths are
allowed to start from the same point. Further, by choosing a suitable topology, the BW can be characterized as a random variable taking values in a
complete separable metric space whose elements are compact sets of paths.
In [16], they gave general convergence criteria allowing either crossing or noncrossing paths, and they verified the criteria for the noncrossing paths case
for coalescing simple random walks. Recently, Ferrari, Fontes, and Wu [14]
verified the same criteria for the noncrossing paths case for a two dimensional
Poisson tree.
The main goal of this paper is to prove weak convergence to the BW for
coalescing nonsimple random walks, which are models with crossing paths.
Technically, the convergence criteria for the crossing paths case are significantly more difficult to verify than the noncrossing paths case. In the noncrossing paths case, the paths form a totally ordered set and one expects
FKG type of positive correlation inequalities to apply; this has been the
main tool in the verification of the convergence criteria for coalescing simple
random walks [16] and a two dimensional Poisson tree [14]. Also, the noncrossing property of paths gives tightness almost for free. For the crossing
23

paths case, tightness needs to be checked separately. Furthermore, correlation inequalities in general no longer apply, and new ideas are needed to
verify a key convergence criterion denoted (B2′ ), as formulated in [16]. For coalescing nonsimple random walks, criterion (B2′ ) turns out to be particularly
difficult to verify (and this has not yet been done). Instead, we formulate
an alternative criterion (E1 ), which we verify along with tightness and the

other convergence criteria. Thus the main contributions of this paper are
contained in Sections 4 and 6 where tightness and criterion (E1 ) are verified.
The new convergence criteria and our approach in verifying them should serve
as a paradigm for establishing the weak convergence of general models with
crossing paths to the Brownian Web. A consequence of the weak convergence
of coalescing nonsimple random walks to the Brownian web is that, for the
dual one-dimensional voter model with initial condition 1’s on the negative
axis and 0’s on the positive axis, the interface region evolves as a standard
Brownian motion after diffusive scaling. This partially recovers and extends
a result of Cox and Durrett [9], which required rather difficult calculations.
Coalescing Random Walks: Let Y , a random variable with distribution
µY , denote the increment of an irreducible aperiodic random walk on Z. We
will always assume E[Y ] = 0, E[Y 2 ] = σ 2 < +∞ throughout this paper,
unless a weaker hypotheses is explicitly stated. For our main result, we will
also need E[|Y |5 ] < +∞. We are interested in the discrete time process of
coalescing random walks with one walker starting from every site on Z ×
Z (first coordinate space, second coordinate time). All walkers have i.i.d.
increments distributed as Y , and two walkers move independently until they
first meet, at which time they coalesce. The path of a random walk is defined
to be the linear interpolation of the random walk’s position at integer times.

Note that for non-simple random walks, two random walk paths can cross
each other many times before they eventually coalesce. If Y were such that
the random walks had period d 6= 1, as in the case of simple random walks
where d = 2, then we would just have d different copies of coalescing random
walks on different space-time sublattices, none of which interacts with the
other copies.
Let X1 (with distribution µ1 ) denote the random realization of such a
collection of coalescing random walk paths on Z × Z, and let Xδ (with distribution µδ ), δ < 1, be X1 rescaled with the usual diffusive scaling of δ/σ in
space, δ 2 in time. The main result of our paper (see Theorem 1.5 below) is
that if E[|Y |5 ] < +∞, then Xδ converges in distribution to the BW as δ → 0.
24

We remark that there are natural interacting particle systems constructed
out of simple random walks, but still with paths crossing each other before
eventual coalescence, for which the methods of this paper should be applicable to prove convergence to the BW. For example, there is a coalescing simple
random walk model on the space time lattice Z × Z, dual to the Stepping
Stone Model with no mutation (see, e.g., [21, 28, 13]), which may be defined
as follows. One walker starts from every site on the lattice Z×Z×{1, · · · , M },
where the first two coordinates are space and time, and the third coordinate
can be regarded as the color. Each walker makes transitions on the spacetime lattice as a simple random walk, and at every transition, a color is

independently chosen from {1, · · · , M } with equal probability. Two walkers make their space-time transitions and choose their corresponding colors
independently until after they meet at the same space-time site and choose
the same color, at which time they coalesce. Projecting the random walk
paths onto the space-time plane Z × Z, and restricting attention to random
walks on the even sublattice (i.e., (x, y) ∈ Z × Z with x + y even), we obtain
a system of coalescing simple random walks with crossing paths. A proof
of convergence for this and other similar models to the BW would require
verification of the hypotheses of Theorem 1.4 below (see also Remark 1.6).
Brownian Web: We recall here Fontes, Isopi, Newman and Ravishankar’s
[15, 16] choice of the metric space in which the Brownian Web takes its values.
Let (R̄2 , ρ) be the completion (or compactification) of R2 under the metric
ρ, where
¯
¯
¯ tanh(x1 ) tanh(x2 ) ¯
¯ ∨ | tanh(t1 ) − tanh(t2 )|. (1.1)
ρ((x1 , t1 ), (x2 , t2 )) = ¯¯
−
1 + |t1 |
1 + |t2 | ¯

R̄2 can be thought of as the image of [−∞, ∞] × [−∞, ∞] under the mapping
¶
µ
tanh(x)
, tanh(t) .
(1.2)
(x, t) Ã (Φ(x, t), Ψ(t)) ≡
1 + |t|

For t0 ∈ [−∞, ∞], let C[t0 ] denote the set of functions f from [t0 , ∞] to
[−∞, ∞] such that Φ(f (t), t) is continuous. Then define
[
Π=
C[t0 ] × {t0 },
(1.3)
t0 ∈[−∞,∞]

where (f, t0 ) ∈ Π represents a path in R̄2 starting at (f (t0 ), t0 ). For(f, t0 ) in
Π, we denote by fˆ the function that extends f to all [−∞, ∞] by setting it

25

equal to f (t0 ) for t < t0 . Then we take
d((f1 , t1 ), (f2 , t2 )) = (sup |Φ(fˆ1 (t), t) − Φ(fˆ2 (t), t)|) ∨ |Ψ(t1 ) − Ψ(t2 )|. (1.4)
t

(Π, d) is a complete separable metric space.
Let now H denote the set of compact subsets of (Π, d), with dH the
induced Hausdorff metric, i.e.,
dH (K1 , K2 ) = sup inf d(g1 , g2 ) ∨ sup inf d(g1 , g2 ).
g2 ∈K2 g1 ∈K1

g1 ∈K1 g2 ∈K2

(1.5)

(H, dH ) is also a complete separable metric space. Let FH denote the Borel
σ-algebra generated by dH .
Lemma 1.1 Assume E[|Y |] < +∞, then for any δ ∈ (0, 1], the closure of
Xδ in (Π, d) is almost surely a compact subset of (Π, d).

Proof. We prove the lemma only for X1 , since the proof for Xδ is identical.
Denote the image of X1 under the mapping (Φ, Ψ) by X1′ . We will show the
equicontinuity of paths in X1′ . Note that by the properties of (Φ, Ψ), this
reduces to showing the equicontinuity of X1′ restricted to any time interval
Ψ(k+1)
[Ψ(k), Ψ(k + 1)] with k ∈ Z, which we denote by X1′ |Ψ(k) . Similarly denote
k+1
the restriction of X1 to the time interval [k, k +1] by X1 |k+1
k . Note that X1 |k
consists of line segments corresponding to random walks jumping from sites
Ψ(k+1)
in Z at time k to sites in Z at time k + 1. If X1′ |Ψ(k) is not equicontinuous,
then there would exist a sequence of random walk jumps from sites xn → −∞
to sites yn > L for some L ∈ R (or from xn → +∞ to yn < L). Since
E[|Y |] < +∞, using Borel-Cantelli, it is easily seen that this is an event
Ψ(k+1)
with probability 0, hence X1′ |Ψ(k) is almost surely equicontinuous, and this
proves the lemma.
Remark 1.1 Note that the closure of Xδ , which from now on we also denote
by Xδ , is obtained from Xδ by adding all the paths of the form (f, t) with

t ∈ δ 2 Z ∪ {+∞, −∞} and f ≡ +∞ or f ≡ −∞. Xδ is then a (H, FH )-valued
random variable.
In [15, 16], the Brownian Web (W̄ with measure µW̄ ) is constructed as a
(H, FH ) valued random variable, with the following characterization theorem.
26

Theorem 1.2 There is an (H, FH )-valued random variable W̄ whose distribution is uniquely determined by the following three properties.
(o) from any deterministic point (x, t) in R2 , there is almost surely a unique
path Wx,t starting from (x, t).
(i) for any deterministic n, (x1 , t1 ), . . . , (xn , tn ), the joint distribution of
Wx1 ,t1 , . . . , Wxn ,tn is that of coalescing Brownian motions (with unit
diffusion constant), and
(ii) for any deterministic, dense countable subset D of R2 , almost surely,
W̄ is the closure in (H, dH ) of {Wx,t : (x, t) ∈ D}.
The (H, FH )-valued random variable W̄ in Theorem 1.2 is called the standard
Brownian Web.
Convergence Criteria and Main Result: In [16], there was also given
a set of general convergence criteria for measures supported on compact sets
of paths which can cross each other. We now state these criteria as four
conditions on our random variables Xδ .
(I1 ) There exist single path valued random variables θδy ∈ Xδ , for y ∈ R2 ,
satisfying: for D a deterministic countable dense subset of R2 , for any de1
m
terministic y 1 , . . . , y m ∈ D, θδy , . . . , θδy converge jointly in distribution as
δ → 0+ to coalescing Brownian motions (with unit diffusion constant) starting at y 1 , . . . , y m .
Let ΛL,T = [−L, L] × [−T, T ] ⊂ R2 . For x0 , t0 ∈ R and u, t > 0, let
R(x0 , t0 ; u, t) denote the rectangle [x0 − u, x0 + u] × [t0 , t0 + t] in R2 . Define
At,u (x0 , t0 ) to be the event (in FH ) that K (in H) contains a path touching both R(x0 , t0 ; u, t) and (at a later time) the left or right boundary of
the bigger rectangle R(x0 , t0 ; 17u, 2t) (the number 17 is chosen to avoid fractions later). Then the following is a tightness condition for {Xδ }: for every
u, L, T ∈ (0, +∞),
(T1 ) g̃(t, u; L, T ) ≡ t−1 lim sup
δ→0+

sup
(x0 ,t0 )∈ΛL,T

µδ (At,u (x0 , t0 )) → 0 as t → 0+

As shown in [16], if (T1 ) is satisfied, one can construct compact sets Gǫ ⊂ H
for each ǫ > 0, such that µδ (Gcǫ ) < ǫ uniformly in δ. Gǫ consists of compact
27

subsets of Π whose image under the map (Φ, Ψ) are equicontinuous with a
modulus of continuity dependent on ǫ.
For K ∈ H a compact set of paths in Π, define the counting variable
Nt0 ,t ([a, b]) for a, b, t0 , t ∈ R, a < b, t > 0 by
Nt0 ,t ([a, b]) = {y ∈ R | ∃ x ∈ [a, b] and a path in K which touches
both (x, t0 ) and (y, t0 + t)}.
(1.6)
Let lt0 (resp. rt0 ) denote the leftmost (resp. rightmost) value in [a, b] with
some path in K touching (lt0 , t0 ) (resp. (rt0 , t0 )). Also define Nt+0 ,t ([a, b])
(resp. Nt−0 ,t ([a, b])) to be the subset of Nt0 ,t ([a, b]) due to paths in K that
touch (lt0 , t0 ) (resp. (rt0 , t0 )). The last two conditions for the convergence of
{Xδ } to the Brownian Web are
(B1′ ) ∀β > 0, lim sup sup sup µδ (|Nt0 ,t ([a − ǫ, a + ǫ])| > 1) → 0 as ǫ → 0+
δ→0+

(B2′ ) ∀β > 0,

t>β t0 ,a∈R

1
lim sup sup sup µδ (Nt0 ,t ([a − ǫ, a + ǫ]) 6= Nt+0 ,t ([a − ǫ, a + ǫ])
ǫ δ→0+ t>β t0 ,a∈R
∪Nt−0 ,t ([a − ǫ, a + ǫ])) → 0 as ǫ → 0+ .

The general convergence theorem of [16] is the following, where we have
replaced our family {Xδ } with δ → 0 by a general sequence {Xn } with
n → +∞:
Theorem 1.3 Let {Xn } be a family of (H, FH ) valued random variables satisfying conditions (I1 ), (T1 ), (B1′ ) and (B2′ ), then Xn converges in distribution
to the standard Brownian Web W̄.
Condition (B1′ ) guarantees that for any subsequential limit X of {Xn }, and
for any deterministic point y ∈ R2 , there is µX almost surely at most one path
starting from y. Together with condition (I1 ), this implies that for a deterministic countable dense set D ⊂ R2 , the distribution of paths in X starting
from finite subsets of D is that of coalescing Brownian motions. Conditions
(B1′ ) and (B2′ ) together imply that for the family of counting random variables
η(t0 , t; a, b) = |Nt0 ,t ([a, b])|, we have E[ηX (t0 , t; a, b)] ≤ E[ηW̄ (t0 , t; a, b)] =
b−a
1+ √
for all t0 , t, a, b ∈ R with t > 0, a < b. By Theorem 4.6 in [16] and
πt
the remark following it, X is then equidistributed with W̄. For the process
28

of coalescing random walks Xδ , we have not yet been able to verify condition
(B2′ ), but an examination of the proof of Theorem 4.6 in [16] shows that we
can also use the dual family of counting random variables
η̂X (t0 , t; a, b) = |{x ∈ (a, b) | ∃ a path in X touching
(1.7)
both R × {t0 } and (x, t0 + t)}|.
By a duality argument [27] (see also [1, 2, 16]), η̂ and η − 1 are equally
distributed for the Brownian Web W̄. We can then replace (B2′ ) by
(E1 ) If X is any subsequential limit of {Xδ }, then ∀t0 , t, a, b ∈ R with t > 0
b−a
.
and a < b, E[η̂X (t0 , t; a, b)] ≤ E[η̂W̄ (t0 , t; a, b)] = √
πt
With this change, we immediately obtain our modified general convergence
theorem,
Theorem 1.4 Let {Xn } be a family of (H, FH ) valued random variables satisfying conditions (I1 ), (T1 ), (B1′ ) and (E1 ), then Xn converges in distribution
to the standard Brownian Web W̄.
The main result of this paper is
Theorem 1.5 If the random walk increment Y satisfies E[|Y |5 ] < +∞, then
{Xδ } satisfy the conditions of Theorem 1.4, and hence converges in distribution to W̄.
Remark 1.6 The main difficulty lies in the verification of the tightness condition (T1 ) and condition (E1 ). Condition (E1 ) in our general convergence
result, Theorem 1.4, may seem strong since it requires an upper bound that is
actually exact, but as we will show in our proof of (E1 ) for {Xδ } in Section 6,
all we need are the Markov property of the random walks and an upper bound
of the type lim supδ↓0 E[η̂Xδ (t0 , t; a, b)] ≤ C for some finite C depending on
t, a, b.
Remark 1.7 Recently, Belhaouari et al. [8] have succeeded in verifying a
version of the tightness criterion (T1 ) in the context of voter model interfaces
under a finite 3 + ǫ moment assumption on Y . This establishes the convergence of Xδ to the Brownian web also under a finite 3+ǫ moment assumption
since the other convergence criteria require either tightness or at most finite
second moment of Y .
29

In Section 2, we list some basic facts about random walks, then in sections
3 to 6, we proceed to verify condition (B1′ ), (T1 ), (I1 ) and (E1 ). In section
7, we present some corollaries for one dimensional coalescing random walks
and the dual non-nearest-neighbor voter models. In particular, we prove
that under the assumption E[|Y |5 ] < +∞, the point process at rescaled
time 1 of coalescing nonsimple random walks starting from (δ/σ)Z at time
0 converges weakly to the point process at time 1 of coalescing Brownian
motions starting from R at time 0. This extends a result of Arratia [1], and
it follows that the density of coalescing
nonsimple random walks starting
√
from Z at time 0 decays as 1/(σ πt), extending a result of Bramson and
Griffeath [5]. Another corollary that follows from the convergence of the
coalescing random walks Xδ to the Brownian Web is that, for the dual voter
−
−
−
model φZt (x) with initial configuration φZ0 (x) ≡ 1 for x ∈ Z− and φZ0 (x) ≡
0 for x ∈ Z+ ∪ {0}, under diffusive scaling, the time evolution of the interface
between the all 0 region and the all 1 region converges in distribution to a
Brownian motion starting from the origin at time 0. This partially recovers
and also extends a result of Cox and Durrett [9], in which they proved that
under the assumption E[|Y |3 ] < +∞, the interface region is of size O(1) as
t → +∞, and under diffusive scaling, the position at rescaled time 1 of the
interface converges in distribution to a standard Gaussian. In Section 7, we
also discuss how our proof can be modified to establish the weak convergence
of continuous time coalescing nonsimple random walks to the Brownian Web.

2

Random Walk Estimates

In this section, we introduce some notation that will be used throughout the
paper, and we list some basic facts about random walks that will be used
in later sections. The results are all standard, but for self-containedness we
include them here.
Given macroscopic space and time coordinates (x, t) ∈ R2 , define their
microscopic counterparts before diffusive scaling by t̃ = tδ −2 and x̃ = xσδ −1 .
Quantities such as ũ, t̃0 are defined from u, t0 similarly. Since µδ and µ1 are
related by diffusive scaling, we will do most of our analysis using µ1 , with
x, t, u, t0 for µδ replaced by x̃, t̃, ũ, t̃0 for µ1 .
Let ξsA denote the state at time s of a system of coalescing random walks
starting from a subset A ⊂ Z × Z. In the special case when A = B × {t0 }
for some B ⊂ Z, we will denote it by ξsB,t0 , and when t0 = 0, we may simply
30

denote it by ξsB , as in the case of B = Z. We will use µ1 (·) to denote the
probability of events for systems of coalescing random walks on Z × Z since
they are marginals of X1 .
Denote the linearly interpolated path of a random walk starting at some
point (x, t0 ) ∈ Z × Z by π x,t0 (s). Denote the event that a random walk
starting at (x, t0 ) stays inside the interval [a, b] containing x up to time t by
x,t0
.
B[a,b],t
Given r ∈ Z, define stopping times
τrx,t0 = inf{n ≥ t0 , n ∈ Z | π x,t0 (n) = r},
0
τrx,t
= inf{n ≥ t0 , n ∈ Z | π x,t0 (n) ≥ r}.
+

(2.1)
(2.2)

When the time coordinate in the superscripts of ξ, π, B, τ, τ+ is 0, we will
suppress it. We will use Px and Ex to denote probability and expectation
for a random walk process starting from x at time 0. Px,y and Ex,y will
correspond to two independent random walks starting at x and y at time 0.
Recall that we always assume the random walk increment Y is distributed
such that the random walk is irreducible and aperiodic with E(Y ) = 0 and
E[Y 2 ] < +∞, unless a different moment condition is explicitly stated.
Lemma 2.1 Let π x , π y be two independent random walks with increment Y
starting at x, y ∈ Z at time 0. Let τx,y be the integer stopping time when the
two walkers first meet, and let l(x, y) = supn∈[0,τx,y ] |π x (n) − π y (n)|. Then
τx,y and l(x, y) are almost surely finite.
Proof.Let π̄ y−x (n) = π y (n) − π x (n). Then π̄ y−x is an irreducible aperiodic
symmetric random walk starting at y − x with increment distributed as µY ∗
µ−Y . The lemma is simply a consequence of the recurrence of π̄ y−x , which
requires less than a finite second moment of Y .
Lemma 2.2 Let π x , π y , τx,y be as in Lemma 2.1. Then Px,y (τx,y > t) ≤
C
√
|x − y| for some constant C independent of t, x and y.
t
Proof. Let π̄ y−x (n) = π y (n) − π x (n) as in the proof of Lemma 2.1. Let
P̄y−x denote probability for this random walk, and let τ̄0y−x denote the first
integer time when π̄ y−x = 0. Then Px,y (τx,y > t) = P̄y−x (τ̄0y−x > t). When
|x − y| = 1, it is a standard fact (see e.g. Proposition 32.4 in [24]) that this
probability is bounded by √Ct . When |x − y| > 1, without loss of generality
31

assume x < y, and regard π x and π y as a subset of the system of coalescing
random walks ξ {x,x+1,...,y} . Then
{x,...,y}

| > 1)
Px,y (τx,y > t) ≤ µ1 (|ξt
y−1
[
C(y − x)
√
.
= µ1 ( {τi,i+1 > t}) ≤ (y − x)P0,1 (τ0,1 > t) ≤
t
i=x
This establishes the lemma.
Lemma 2.3 Let u > 0 and t > 0 be fixed, and let π(s) = π 0,0 (s) be a random
0
walk starting from the origin at time 0. Let ũ, t̃ and the event B[−ũ,ũ],
be
t̃
defined as at the beginning of this section (note that they depend on δ), and
0
0
let (B[−ũ,ũ],
)c be the complement of B[−ũ,ũ],
. If Bs is a standard Brownian
t̃
t̃
motion starting from 0, then
u2

0
c
− 2t
.
0 < lim+ P0 ((B[−ũ,ũ],
t̃ ) ) = P( sup |Bs | > u) < 4e
δ→0

s∈[0,t]

Proof. The limit follows from Donsker’s invariance principle [11] for random
walks. The first inequality is trivial, and the second inequality follows from a
well-known computation for Brownian motion using the reflection principle.
Lemma 2.4 Let u, t, ũ, t̃ be as before. Let π x , π y and τx,y be as in Lemmas
2.1−2.2 with x < y. Let τx,y,ũ+ be the first integer time when π x (n)−π y (n) ≥
ũ. If E[|Y |3 ] < +∞, and δ is sufficiently small, we then have
Px,y (τx,y,ũ+ < (τx,y ∧ t̃)) < C(t, u)δ
for some constant C(t, u) depending only on t and u.
Proof. Let z = x − y < 0. Note that x, y, z are fixed while ũ, t̃ → +∞
as δ → 0. For the difference of the two walks π̄ z (n) = π x (n) − π y (n), we
still denote the first integer time when π̄ z = 0 by τ̄0z , and the first integer
time when π̄ z ≥ ũ by τ̄ũz+ . Note we are using the bar ¯· to emphasize the fact
that we are studying the symmetrized random walks. The inequality then
becomes
P̄z (τ̄ũz+ < (τ̄0z ∧ t̃)) < C(t, u)δ.
32

(2.3)

First we will prove that, for δ sufficiently small,
P̄w (τ̄ũw+ < (τ̄0w ∧ t̃)) < C ′ (t, u)|w|δ.

(2.4)

By the strong Markov property,
P̄w (τ̄0w > t̃)
+∞
X
k,τ̄ w
≥
P̄w (τ̄ũw+ < (τ̄0w ∧ t̃), π̄ w (τ̄ũw+ ) = k, B[k−ũũ+,k+ ũ ],t̃ )
2

k=⌈ũ⌉

=

⌊t̃⌋
+∞ X
X

k
P̄w (τ̄ũw+ = n, n < τ̄0w , π̄ w (n) = k)P̄k (B[k−
)
ũ
,k+ ũ ],t̃−n

⌊t̃⌋
+∞ X
X

k
)
P̄w (τ̄ũw+ = n, n < τ̄0w , π̄ w (n) = k)P̄k (B[k−
ũ
,k+ ũ ],t̃

2

k=⌈ũ⌉ n=0

≥

2

2

k=⌈ũ⌉ n=0

2

2

≥ C ′′ (t, u)P̄w (τ̄ũw+ < (τ̄0w ∧ t̃)).
If δ is sufficiently small, the last inequality is valid by Lemma 2.3. Also by
Lemma 2.2,
C
C|w|
P̄w (τ̄0w > t̃) < √ |w| = √ δ,
t
t̃
together they give (2.4).
To show (2.3), we condition on the first integer time when π̄ z (n) ≥ 0,
which we denote by τ̄0z+ . Then by the strong Markov property and (2.4),
P̄z (τ̄ũz+ < (τ̄0z ∧ t̃))
=

⌊t̃⌋
+∞ X
X

P̄z [τ̄0z+ = n, π̄ z (n) = w]P̄w [τ̄ũw+ < (τ̄0w ∧ (t̃ − n))]

⌊t̃⌋
+∞ X
X

P̄z [τ̄0z+ = n, π̄ z (n) = w] C ′ (t, u)|w| δ

w=1 n=0

<

w=1 n=0

< C ′ (t, u) δ Ēz [π̄ z (τ̄0z+ )] < C(t, u)δ.
The last inequality follows from our assumption E[|Y |3 ] < +∞ and the
following two lemmas.
33

Lemma 2.5 Let π x be a random walk with increment Y starting from x <
0 at time 0. If E[Y 2 ] < +∞, then the overshoot π x (τ0x+ ) has a limiting
distribution as x → −∞. In terms of the ladder variable Z = π 0 (τ10+ ),
lim P[π x (τ0x+ ) = k] =

x→−∞

P[Z ≥ k + 1]
.
E[Z]

Proof.This is a standard fact from renewal theory, see e.g. Proposition 24.7
in [24].
Lemma 2.6 Let π x , Y and Z be as in the previous lemma. If E[|Y |r+2 ] <
+∞ for some r > 0, then [π x (τ0x+ )]r is uniformly integrable in x ∈ Z− , and
+∞

¤
£
lim E [π x (τ0x+ )]r =

x→−∞

1 X r
k P[Z ≥ k + 1] < +∞.
E[Z] k=1

Proof. Note that if we let γ x denote the random walk starting from x < 0 at
time 0 with increment distributed as Z, then γ x simply records the successive
maxima of the random walk π x , and so the overshoots π x (τ0x+ ) and γ x (τ0x+ )
are equally distributed. By a last passage decomposition for γ x ,
P[γ

x

(τ0x+ )

= k] =

−1
X
i=x

Gγ (x, i)P[Z = k − i] ≤ P[Z ≥ k + 1],

where Gγ (x, i) is the probability γ x will ever visit i. Since E[|Y |r+2 ] < +∞
r+1
implies
(see e.g. problem 6 in Chapter IV of [24]), we have
£ x xE[Zr ¤ ] <
P+∞
+∞ r
E [π (τ0+ )] ≤ k=1 k P[Z ≥ k + 1] < +∞, giving uniform integrability.
The rest then follows from Lemma 2.5 and dominated convergence.
Lemma 2.7 Let ξnZ be a system of coalescing random walks starting from
every site on Z at time 0, whose random walk increments are distributed as
Y with E[Y 2 ] < +∞. Then pn ≡ µ1 (0 ∈ ξnZ ) ≤ √Cn for some constant C
independent of the time n.
Remark 2.1 The proof we present here is an adaptation of the argument
used by Bramson and Griffeath [5] to establish similar upper bounds for cond
tinuous time coalescing simple random walks
√ in Z , d ≥ 2. In Corollary 7.1
below, we will prove that in fact pn ∼ 1/(σ πn) as n → +∞ under a stronger
moment assumption.
34

Proof. Let BM = [0, M − 1] ∩ Z, and let en (BM ) = E[|ξnZ ∩ BM |]. By
translation invariance, en (BM ) = pn M , and
X
X
en (BM ) ≤
E[|ξnBM +kM ∩ BM |] =
E[|ξnBM ∩ (BM + kM )|] = E[|ξnBM |].
k∈Z

k∈Z

Since M − |ξnBM | is at least as large as the number of nearest neighbor pairs
in BM that have coalesced by time n, we may take expectation and apply
Lemma 2.2 to obtain
C
C
E[|ξnBM |] ≤ M −(M −1)µ1 (|ξn{0,1} | = 1) ≤ M −(M −1)(1− √ ) < 1+M √ .
n
n
√
Therefore pn < 1/M +√C/ n. Since M can be arbitrarily large for any fixed
n, we obtain pn ≤ C/ n.
Lemma 2.8 For any A ⊂ Z, let ξnA be a system of discrete time coalescing
random walks on Z starting at time 0 with one walker at every site in A,
where all the random walks have increments distributed as some arbitrary
Z-valued random variable Y . Then for any pair of disjoint sets B, C ⊂ Z,
and for any time n ≥ 0,
P(ξnA ∩ B 6= ∅, ξnA ∩ C 6= ∅) ≤ P(ξnA ∩ B 6= ∅)P(ξnA ∩ C 6= ∅).

(2.5)

In particular, if x, y are any two distinct sites in Z, we have
µ1 (x ∈ ξnZ , y ∈ ξnZ ) ≤ µ1 (x ∈ ξnZ )µ1 (y ∈ ξnZ ).

(2.6)

Proof. The continuous time version of this lemma is due to Arratia (see
Lemma 1 in [3]). Our proof for the discrete time case is an adaptation of
Arratia’s proof for the continuous time case. Arratia’s proof uses a theorem of
Harris [18], which breaks down for discrete time because there are transitions
between states that are not comparable to each other. However, this can be
easily remedied by using an induction argument.
We can assume A, B, C are all finite sets, since otherwise we can approximate by finite sets, and the relevant probabilities will all converge. The
main tool in the proof is the duality between coalescing random walks and
voter models. For any pair of finite disjoint sets B, C ⊂ Z, let φB,C
n , with
B,C
distribution νn , be a discrete time multitype voter model on Z defined as
follows. The state space is X = {−1, 0, 1}Z with product topology. At time
35

0, φB,C
(x) = 0 if x ∈ (B ∪ C)c ; φB,C
(x) = 1 if x ∈ B; φB,C
(x) = −1 if x ∈ C.
0
0
0
B,C
B,C
B,C
At time n ≥ 1, we update φn by setting φn (x) = φn−1 (x + Yx,n ), where
{Yx,n }x∈Z,n∈N are i.i.d. integer-valued random variables distributed as Y .
Let EA+ ⊂ X (resp., EA− ⊂ X) be the event that some site in A is assigned
the value +1 (resp., −1). Then by the duality between voter models and
coalescing random walks (see, e.g., [22]), P(ξnA ∩ B 6= ∅) = νnB,C (EA+ ), P(ξnA ∩
C 6= ∅) = νnB,C (EA− ), and P(ξnA ∩ B 6= ∅, ξnA ∩ C 6= ∅) = νnB,C (EA+ ∩ EA− ). The
correlation inequality (2.5) then becomes
νnB,C (EA+ ∩ EA− ) ≤ νnB,C (EA+ )νnB,C (EA− ).

(2.7)

We can define a partial order on the state space X by setting η ≤ ζ ∈ X
whenever η(x) ≤ ζ(x) for all x ∈ Zd .A function f : X → R is called increasing (resp., decreasing) if for any η ≤ ζ, f (η) ≤ f (ζ) (resp., f (η) ≥ f (ζ)). An
event E is called increasing (resp., decreasing) if 1E is an increasing (resp.,
decreasing) function. Clearly, for finite A, 1E + is an increasing function and
A
1E − is a decreasing function. Inequality (2.7) will follow if we show that
A
νnB,C has the FKG Rproperty (see,R e.g, [17,R22]), i.e., for any two increasing
functions f and g, f g dνnB,C ≥ f dνnB,C g dνnB,C .
We prove this by induction. For any pair of finite disjoint sets B, C ⊂ Z,
B,C
ν0 has the FKG property because the measure is concentrated at a single
configuration. Observe that ν1B,C is a product measure and therefore also
has the FKG property (this is the consequence of a simple special case of
the main result of [17]). We proceed to the induction step, which is a fairly
standard argument [19]. Assume that for all disjoint finite sets B and C,
and for all 0 ≤ k ≤ n − 1, νkB,C has the FKG property. Let us denote the
B,C
collection of sites in Z where φB,C
n−1 (x) = 1 by Bn−1 , and where φn−1 (x) = −1
by Cn−1 . Then for any two increasing functions f and g, conditioning on
φB,C
n−1 , we have by the Markov property,
Z
Z Z
B
,C
B,C
B,C
f g dνn =
f g dν1 n−1 n−1 dνn−1
Z
Z Z
B
,C
Bn−1 ,Cn−1
B,C
g dν1 n−1 n−1 dνn−1
≥
f dν1
Z Z
Z Z
Bn−1 ,Cn−1
B
,C
B,C
B,C
≥
f dν1
dνn−1
g dν1 n−1 n−1 dνn−1
Z
Z
B,C
=
f dνn
g dνnB,C ,
36

B

,C

B,C
where we have used the FKG property for both νn−1
and for ν1 n−1 n−1 , and
R
R
B
,C
B
,C
the observation that the conditional expectations f dν1 n−1 n−1, gdν1 n−1 n−1
B,C
conditioned on φB,C
also has
n−1 are still increasing functions. Therefore νn
the FKG property. This concludes the induction proof, and establishes the
lemma.

Remark 2.2 Lemma 2.8 is also valid for random walks in Zd by the same
argument.

3

Verification of condition (B1′ )

Let’s fix t0 , a ∈ R, β > 0, t > β, ǫ > 0. Also fix a δ and let t̃0 , t̃, ã and ǫ̃ be
defined from t0 , t, a and ǫ by diffusive scaling. Then we have
µδ (|Nt0 ,t ([a − ǫ, a + ǫ])| > 1) = µ1 (|Nt̃0 ,t̃ ([ã − ǫ̃, ã + ǫ̃])| > 1).
If t̃0 = n0 ∈ Z, then the contribution to N is all due to walkers starting from
[ã − ǫ̃, ã + ǫ̃] ∩ Z at time n0 . Thus we have
[ã−ǫ̃,ã+ǫ̃]∩Z,n0

µ1 (|Nn0 ,t̃ ([ã − ǫ̃, ã + ǫ̃])| > 1) = µ1 (|ξn0 +t̃
≤

⌊ã+ǫ̃⌋−1

X

{i,i+1},n0

µ1 (|ξn0 +t̃

i=⌈ã−ǫ̃⌉
{0,1},0

≤ 2ǫ̃µ1 (|ξt̃

| > 1)

| > 1)

2Cσǫ
2Cσǫ
C
| > 1) ≤ 2ǫ̃ √ = √ < √
β
t
t̃

(3.1)

The first inequality follows from the observation that if the collection of
walkers starting from [ã − ǫ̃, ã + ǫ̃] ∩ Z at n0 has not coalesced into a single
walker by n0 + t̃, then there is at least one adjacent pair of such walkers which
has not coalesced by n0 + t̃. The next inequality follows from Lemma 2.2.
Now suppose t̃0 ∈ (n0 , n0 + 1) for some n0 ∈ Z. Note that a walker’s
path can only cross [ã − ǫ̃, ã + ǫ̃] × {t̃0 } due to the increment at time n0 .
After the increment, at time n0 + 1, it will either land in [ã − 2ǫ̃, ã + 2ǫ̃], or
else outside that interval. In the first case, the contribution of the walker’s
[ã−2ǫ̃,ã+2ǫ̃]∩Z,n0 +1
path to N is included in ξt̃0 +t̃
, the probability of which by our
4Cσǫ
previous argument is bounded by √t times a prefactor which approaches 1
as δ → 0. In the second case, either a walker in (−∞, ã + ǫ̃] jumps to the
37

right of ã + 2ǫ̃, or a walker in [ã − ǫ̃, +∞) jumps to the left of ã − 2ǫ̃, the
probability of which is bounded by
+∞
X

x=⌈ã−ǫ̃⌉

≤
≤

+∞
X

k=0
+∞
X
k=0

P(Y ≤ ã − 2ǫ̃ − x) +

P(|Y | ≥ k + ǫ̃) ≤

+∞
X
k=0

⌊ã+ǫ̃⌋

X

x=−∞

P(Y ≥ ã + 2ǫ̃ − x)

E[Y 2 , |Y | ≥ k + ǫ̃]
(k + ǫ̃)2

2E[Y 2 , |Y | ≥ ǫ̃]
4σ 2
4σδ
E[Y 2 , |Y | ≥ ǫ̃]
≤
≤
=
.
2
(k + ǫ̃)
ǫ̃
ǫ̃
ǫ

(3.2)

The next to last inequality in (3.2) is valid if we take δ to be sufficiently
small. The bounds in (3.1) and (3.2) are independent of t0 , t > β and a.
Taking the supremum over t > β, t0 and a, and letting δ → 0+ , we obtain
(B1′ ).
Corollary 3.1 Assume X (with distribution µ) is a subsequential limit of
Xδ , then for any deterministic point y ∈ R2 , X has almost surely at most
one path starting from y.
Proof. It was shown in the proof of Theorem 5.3 in [16] that (B1′ ) implies
(B1′′ ) ∀β > 0, sup sup µ(|Nt0 ,t ([a − ǫ, a + ǫ])| > 1) → 0 as ǫ → 0+
t>β t0 ,a∈R

and the corollary then follows.
Remark 3.1 Note that if ZδAδ is the process of coalescing random walks
starting from a subset Aδ of the rescaled lattice, and ZδAδ converges in distribution to a limit Z, then by the same argument as above, for any deterministic
point y ∈ R2 , Z has almost surely at most one path starting from y.

4

Verification of tightness condition (T1)

Define A+
t,u (x0 , t0 ) to be the event that K contains a path touching both
R(x0 , t0 ; u, t) and (at a later time) the right boundary of the bigger rectangle
R(x0 , t0 ; 17u, 2t), and similarly define the event A−
t,u (x0 , t0 ) corresponding to
38

the left boundary of the bigger rectangle. Then A = (A+ ∪ A− ), and writing
(T1 ) in terms of µ1 , we argue that it is sufficient to prove
(T1+ )

g̃(t, u; L, T ) ≡ t−1 lim sup µ1 (A+
(0, 0)) → 0 as t → 0+ .
t̃,ũ
δ→0+

The sup over x0 , t0 has been safely omitted because µ1 is invariant under
translation by integer units of space and time. When x̃0 , t̃0 ∈
/ Z, we can
bound the probability from above by using larger rectangles with vertices
in Z × Z and base centered at (0, 0). Since the argument establishing the
analogous tightness condition (T1− ) for the event A− is identical to that for
(T1+ ), (T1 ) follows from (T1+ ). In the ensuing discussions, we will simply write
+
+
A+
t,u (0, 0) as At,u or just A , and R(0, 0; u, t) as R(u, t).
Before we prove (T1+ ), and hence (T1 ), we introduce some simplifying
notation. Denote random walks starting at time 0 from x1 = ⌈3ũ⌉, x2 =
⌈7ũ⌉, x3 = ⌈11ũ⌉, x4 = ⌈15ũ⌉ by π1 , π2 , π3 , π4 . Denote the event that πi ,
(i = 1, 2, 3, 4) stays within a distance ũ of xi up to time 2t̃ by Bi (see
Figure 1). For a random walk starting from (x, m) ∈ R(ũ, t̃), denote the
integer stopping times that the walker first exceeds 5ũ, 9ũ, 13ũ and 17ũ by
τ1x,m , τ2x,m , τ3x,m and τ4x,m . We also define τ0x,m = m, and τ5x,m = 2t̃. Denote
the event that π x,m does not coalesce with πi before time 2t̃ by Ci (x, m).
As we shall see, the reason for choosing four paths πi is because each path
contributes a factor of δ to our estimate of the µ1 probability in (T1+ ), and
an overall factor of δ 4 is needed to outweigh the O(δ −3 ) number of lattice
points in the rectangle R(ũ, t̃) from where a random walk can start. We are
now ready to prove (T1+ ).
Proof of (T1+ ). First we can assume t̃ ∈ Z, since we can always replace
t̃ by ⌈t̃⌉ which only enlarges the event A+ . The contribution to the event
A+ is either due to random walk paths that originate from within R(ũ, t̃), or
paths that cross R(ũ, t̃) without landing inside it after the crossing. Denote
the latter event by D(ũ, t̃). Then
µ1 (A+
)
t̃,ũ

≤ µ1 (

4
[

i=1
4
\

+ µ1 (

i=1

Bic ) + µ1 (D(ũ, t̃))
Bi ; ∃(x, m) ∈ R(ũ, t̃) s.t.

4
\

Ci (x, m), τ4x,m < 2t̃). (4.1)

i=1

By Lemma 2.3, the first term on the right hand side of (4.1) is of order
o(t) after taking the limit δ → 0. For the second event in (4.1) to occur,
39

π1

2t̃

π2

π3

π4
τ4
τ3

τ2

t̃
τ1
π x,m
0
0

ũ

2ũ 3ũ

4ũ

5ũ

6ũ 7ũ

8ũ 9ũ

10ũ 11ũ 12ũ 13ũ 14ũ 15ũ 16ũ 17ũ

Figure 1: The random walks π1 , π2 , π3 and π4 start from 3ũ, 7ũ, 11ũ and 15ũ
at time 0 and each stays within a distance of ũ. The random walk π x,m starts
from (x, m) inside the rectangle R(0, 0; ũ, t̃) and exits the right boundary of
the rectangle R(0, 0; 17ũ, 2t̃) at time τ4 without coalescing with π1 , π2 , π3 and
π4 on the way.

either a walker at a site in (−∞, −ũ] × {n} jumps to a site in [ũ, +∞)
in one step, or a walker in [ũ, +∞) × {n} jumps to a site in (−∞, −ũ]
in one step for some n ∈ [0, t̃ − 1] ∩ Z. Denote the event just described
Pt̃−1
by D′ (ũ, n), then µ1 (D(ũ, t̃)) ≤ n=0
µ1 (D′ (ũ, n)). From this we see that
repeating the calculations in (3.2) for (B1′ ) under the assumption E[|Y |3 ] <
+∞ will guarantee that µ1 (D(ũ, t̃)) → 0 as δ → 0+ .
To estimate the third term in (4.1) (see Figure 1 for an illustration of the
event), we first treat the case of a fixed (x, m) ∈ R(ũ, t̃). Suppressing (x, m)
from π x,m , Ci (x, m) and τix,m , we have
µ1 {for (x, m) fixed,

4
\

i=1

Bi ,

4
\

Ci , τ4 < 2t̃}

i=1

1
1
1
≤ µ1 {π(τ1 ) > 5 ũ or π(τ2 ) > 9 ũ or π(τ3 ) > 13 ũ}
2
2
2
\
\
1
1
1
+ µ1 {π(τ1 ) ≤ 5 ũ, π(τ2 ) ≤ 9 ũ, π(τ3 ) ≤ 13 ũ, τ4 < 2t̃, Bi , Ci }. (4.2)
2
2
2
The first part is bounded by
3 sup Px [π x (τ0+ ) >
x∈Z−

ũ
ũ
2
] ≤ 3 ( )3 sup Ex [(π x (τ0+ ))3 , π x (τ0+ ) > ]
2
ũ x∈Z−
2
24
(4.3)
≤ 3 3 δ 3 ω(δ),
uσ
40

where ω(δ) → 0 as δ → 0. The last inequality is due to the uniform integrability of the third moment of the overshoot distribution, which follows from
our assumption E[Y 5 ] < +∞ and Lemma 2.6.
For the second µ1 probability in (4.2), denote the event that none of the
conditions listed are violated by time t by Gt . If τ1 > t, we interpret an
inequality like π(τ1 ) ≤ 5 12 ũ as not having been violated by time t. Gt is then
a nested family of events, and the second probability in (4.2) becomes
µ1 (G2t̃ ) = µ1 (Gτ5 ) = µ1 (Gm )

5
Y

k=1

µ1 (Gτk |Gτk−1 ) <

4
Y

k=1

µ1 (Gτk |Gτk−1 ),

since Gτk ⊂ Gτk−1 . Denote the history of the random walks π x,m , π 1 , π 2 , π 3
and π 4 up to time t by Πt , and denote expectation with respect to the
conditional distribution of Πt conditioned on the event Gt by Et . Then for
k = 1, 2, 3, 4,
£
¤
µ1 (Gτk |Gτk−1 ) = Eτk−1 µ1 (Gτk |Πτk−1 ∈ Gτk−1 ) ,

where the µ1 probability on the right hand side is conditioned on a given
realization of Πτk−1 ∈ Gτk−1 , which is a positive probability event. For any
Πτk−1 ∈ Gτk−1 , we have by the strong Markov property that

µ1 (Gτk |Πτk−1 ∈ Gτk−1 ) = µ1 [Gτk |π x,m (τk−1 ), π i (τk−1 ), i = 1, 2, 3, 4]
≤ C(t, u)δ,
(4.4)
where the inequality follows from Lemma 2.4 for δ sufficiently small. Thus
µ1 (Gτk |Gτk−1 ) ≤ C(t, u)δ, and µ1 (G2t̃ ) ≤ C 4 (t, u)δ 4 . We then have
µ1 (

4
\

i=1

≤

X

Bi ; ∃(x, m) ∈ R(ũ, t̃), s.t.
X

x∈[−ũ,ũ]∩Z m∈[0,t̃]∩Z

≤ [

4
\

Ci (x, m), τ4x,m < 2t̃)

i=1

4
4
\
\
£
¤
µ1 for (x, m) fixed, Bi ,
Ci , τ4x,m < 2t̃
i=1

24 3
δ ω(δ) + C 4 (t, u)δ 4 ] 2ũ t̃ = ω ′ (δ),
u3 σ 3

i=1

(4.5)

where 2ũt̃ = O(δ −3 ) and hence ω ′ (δ) → 0 as δ → 0. Thus the last two terms
in (4.1) go to 0 as δ → 0, and the first term is of order o(t) after taking the
limit δ → 0. Together they give (T1+ ).
41

Remark 4.1 The only place in this paper where we need the assumption
E[|Y |5 ] < +∞ is in (4.3). We need E[|Y |3 ] < +∞ to estimate µ1 (D(ũ, t̃)) in
(4.1), and to apply Lemma 2.4 in (4.4). A finite third moment is the minimal
moment condition for the convergence of Xδ . For any ǫ > 0, there are choices
of Y satisfying E[|Y |3−ǫ ] < +∞, but with µ1 (D(ũ, t̃)) → 1 as δ → 0 for all
t > 0, which implies {Xδ } is not tight.

5

Verification of condition (I1)

Our verification of (I1 ) is essentially a careful application of Donsker’s invariance principle and the Continuous Mapping Theorem for weak convergence.
We define three sets of random walks: {πδi }1≤i≤m , a family of m independent
i
random walks on the rescaled lattice (δ/σ)Z × δ 2 Z; {πδ,f
}1≤i≤m , the family
i
of m coalescing random walks constructed from {πδ } by applying a mapping
i
f to {πδi }; and {πδ,g
}1≤i≤m , an auxiliary family of m coalescing walks constructed by applying a mapping g to {πδi } such that two walks coalesce as
i
soon as their paths intersect (note that random walks in {πδ,g
} coalesce earlier
i
than they do in {πδ,f }). By Donsker’s invariance principle, {πδi } will converge
weakly to a family of independent Brownian motions {B i }1≤i≤m . As we will
see, the mapping g is almost surely continuous with respect to {B i }, and
{Bgi }1≤i≤m is distributed as coalescing Brownian motions. Therefore by the
i
Continuous Mapping Theorem for weak convergence, {πδ,g
} converges weakly
i
i
to the coalescing Brownian motions {Bg }. Finally to show that {πδ,f
} also
i
converges weakly to {Bg }, we will prove that the distance between the two
i
i
versions of coalescing walks {πδ,f
} and {πδ,g
} converges to 0 in probability.
We start with some notation. Let D be any deterministic countable dense
subset of R2 . Let y 1 = (x1 , t1 ), . . . , y m = (xm , tm ) ∈ D be fixed, and let
B1 , ..., B m be independent Brownian motions starting from y 1 , ..., y m . For a
fixed δ, denote ⌈ỹ i ⌉ = (⌈x̃i ⌉, ⌈t̃i ⌉), where x̃i = σδ −1 xi and t̃ = δ −2 ti as defined
in Section 2, and let yδi denote ⌈ỹ i ⌉’s space-time position after diffusive scaling
on the rescaled lattice (δ/σ)Z × δ 2 Z. Let π̃ i (i = 1, · · · , m) be independent
random walks in the Z × Z lattice starting from ⌈ỹ i ⌉. We regard (B 1 , ..., B m ),
and (π̃ 1 , ..., π̃ m ) as random variables in the product metric space (Πm , d∗m ),
where
d∗m [(ξ 1 , . . . , ξ m ), (ζ 1 , . . . , ζ m )] = max d(ξ i , ζ i )
1≤i≤m

(5.1)

and d is defined in (1.4); thus d∗m gives the product topology on Πm . We
42

will also need the metric
¯ 1 , t1 ), (f2 , t2 )) = sup |fˆ1 (t) − fˆ2 (t)| ∨ |t1 − t2 |
d((f

(5.2)

t

and d¯∗m is defined in a similar way as d∗m .
We now define a mapping g from (Πm , d∗m ) to (Πm , d∗m ) that constructs
coalescing paths from independent paths. The construction is such that when
two paths first intersect, the path with the higher index will be replaced by
the path with the lower index after the time of intersection. This procedure
is then iterated until no more intersections take place. To be explicit, we
give the following algorithmic construction.
Let (ξ 1 , . . . , ξ m ) ∈ Πm , and start with equivalence relations on the set
{1, . . . , m} by letting i ≁ j ∀ i 6= j. Define a one step iteration Γ on
(ξ 1 , . . . , ξ m ) and the equivalence relations by
τg = inf{t ∈ R | ∃ i, j ∈ {1, . . . , m}, i ≁ j, ξ i (t) = ξ j (t)},
i∗ = min{j | 1 ≤ j ≤ m, ξ i (τg ) = ξ j (τg )},
i

Γξ (t) =

½

ξ i (t) if t ≤ τg ,
∗
ξ i (t) if t > τg ,

(5.3)
(5.4)
(5.5)

and update equivalence relations by assigning i ∼ i∗ . Iterate the mapping
Γ, and label the successive intersection times τg by τgk . Then the iteration
stops when τgk = +∞ for some k ∈ {1, 2, . . . , m}, i.e., either there is no
more intersection among the different equivalence classes of paths, or all the
paths have coalesced and formed a single equivalence class. Denote the final
collection of paths by g(ξ 1 , . . . , ξ m ) = (ξg1 , . . . , ξgm ). Then it’s clear by the
strong Markov property, that (Bg1 , . . . , Bgm ) has the distribution of coalescing
Brownian motions. However (π̃g1 , . . . , π̃gm ) is not distributed as coalescing
random walks, because for nonsimple random walks, paths can cross before
the random walks actually coalesce (by being at the same space-time lattice
site).
To construct coalescing random walk paths from independent random
walks in Z × Z, we define another mapping f from (Πm , d∗m ) to (Πm , d∗m )
in a similar way as we defined g, except in (5.3)–(5.5), the time of first
intersection τg is replaced by the time of first coincidence on the unscaled
lattice,
τf = min{ n ∈ Z | ∃ i, j ∈ {1, . . . , m}, i ≁ j, ξ i (n) = ξ j (n)}.
43

(5.6)

We will label the successive coincidence times by τfk . Also denote f (ξ 1 , · · · , ξ m )
by (ξf1 , · · · , ξfm ). It is then clear that (π̃f1 , . . . , π̃fm ) is distributed as coalescing
random walks starting from (⌈ỹ 1 ⌉, . . . , ⌈ỹ m ⌉) in the unscaled lattice Z × Z.
We shall denote the diffusively rescaled versions of (π̃ 1 , . . . , π̃ m ), (π̃g1 , . . . , π̃gm )
1
m
1
m
and (π̃f1 , . . . , π̃fm ) by (πδ1 , . . . , πδm ), (πδ,g
, . . . , πδ,g
) and (πδ,f
, . . . , πδ,f
). We need
the following two lemmas to prove (I1 ).
Lemma 5.1 Let (B 1 , ..., B m ) be m independent Brownian motions starting
from (y 1 , ..., y m ), and let (πδ1 , ..., πδm ) be independent random walks starting
1
m
from (yδ1 , ..., yδm ) in the rescaled lattice as defined before. Then (πδ,g
, ..., πδ,g
)
1
m
+
converges in distribution to (Bg , ..., Bg ) as δ → 0 .
Proof. Clearly (yδ1 , . . . , yδm ) converge to (y 1 , . . . , y m ). By Donsker’s invariance principle [11], (πδ1 , ..., πδm ) converges weakly to (B 1 , ..., B m ) as δ → 0+ .
From standard properties of Brownian motion, it is easy to see that the
mapping g is almost surely continuous with respect to (B 1 , ..., B m ). Also
1
m
note that (πδ,g
, ..., πδ,g
) is the same as applying g to (πδ1 , ..., πδm ), therefore by
the Continuous Mapping Theorem for weak convergence (see, e.g., Section
1
m
8.1 of [12]), (πδ,g
, ..., πδ,g
) converges in distribution to (Bg1 , ..., Bgm ) as δ → 0+ .
1
m
1
m
Lemma 5.2 ∀ ǫ > 0, P{d∗m [(πδ,f
, . . . , πδ,f
), (πδ,g
, . . . , πδ,g
)] ≥ ǫ} → 0 as
+
δ→0 .

Proof. From the definition of d and d¯ in (1.4) and (5.2), it is clear that
¯ 1 , t1 ), (f2 , t2 )) for any (f1 , t1 ), (f2 , t2 ) ∈ Π. Therefore
d((f1 , t1 ), (f2 , t2 )) ≤ d((f
it is sufficient to prove the lemma with d∗m replaced by d¯∗m . In terms of
random walks in Z × Z, the lemma can be stated as
∀ ǫ > 0, P{d¯∗m [(π̃f1 , . . . , π̃fm ), (π̃g1 , . . . , π̃gm )] ≥ ǫ̃} → 0 as δ → 0+ .

(5.7)

We first prove (5.7) for m = 2. Note that for m = 2, π̃f1 = π̃g1 = π̃ 1 ,
hence d¯∗2 [(π̃f1 , π̃f2 ), (π̃g1 , π̃g2 )] = d¯(π̃f2 , π̃g2 ). Let T̃g1,2 denote the first integer
time when π̃ 1 and π̃ 2 coincide or interchange relative ordering, and let T̃f1,2
denote the first integer time when the two walks coincide. Also let l(0, n)
denote the maximum distance over all time between two coalescing random
walks starting at 0 and n at time 0. Then by the strong Markov property,
and conditioning at time T̃g1,2 ,
�